Skip to content
LegalGita
How it works

How we make the system reliable and trustworthy.

Specialists in parallel, a shared registry of authorities, ranking in code, and checks that every citation, quotation and figure must pass before you see it.

01 Orchestration

One question, many specialists, one answer.

A planner reads the question and the documents in play, then sends specialists out in parallel. They share one context graph of what each has found, and one registry of authorities taken from the official text, so every specialist always reasons from the official law.

Planner
Researcher A
  • Bharat Aluminium (BALCO)
  • Vijay Karia
Researcher B
  • Campos Brothers
  • Venture Global
Statutes researcher
  • Section 48, Arbitration Act
Shared authority registryEmpty
Synthesiser writes one answer
02 The legal researcher

A research funnel that reads, not skims.

Statutes are grounded once, before any search. Then the researcher plans its searches, ranks what comes back in code and builds a context graph of cases and issues. An assessor reads the relevant sources in full and re-weighs the graph, and the answer is written from the final graph.

From the question to the cases that hold.

1 of 7
Ground
The law first

Before any search, the provisions in issue are read from India Code and the e-Gazette, and a repealed IPC, CrPC or Evidence Act section is paired with its successor.

2 of 7
Plan
Searches for each issue

Across the case-law index, the courts' own sites, the reputed legal press and the open web. One search always stays open, so the plan can find what nobody thought to name.

3 of 7
Rank
Ranked in code

Results are pooled by agreement across searches and ranked on five signals. One judgment behind four links takes one slot, and the court's own copy wins over its mirrors.

4 of 7
Graph
A context graph

Every ranked case is linked to the issues it bears on, and to the cases it follows or overrules.

5 of 7
Read
Read in full, then re-weighed

An assessor starts from the context graph and reads the relevant sources in full. It rescores each one and changes its weight in the graph.

6 of 7
Verify
Each case still stands

Every case the answer relies on is found on the court's own site, checked to exist as cited, and checked for what happened to it later.

7 of 7
Cite
Written from the final graph

Within each issue, to see which case leads and which is doubtful, and across issues, to find the case that anchors several and the one that cuts both ways. It cites only what it read.

Ranked in code, not by the model.

Five signals, computed from the text and dates alone. No model call decides what gets read.

  • Agreement

    How many independent searches surfaced the source

  • Authority

    Where the court sits in the hierarchy, and what commentary reports on

  • Recency

    Newer authority weighs more, halving over four years

  • Source

    The court's own record above a transcription, above commentary

  • Statutory anchor

    Whether the provision in issue actually appears

03 Citation integrity

The answer cannot cite what was not read.

Specialists never type a case name. They cite by handle, and a handle only exists for a source that was actually retrieved. When the answer is written, handles become real names; a handle that matches nothing takes its sentence with it.

Retrieved sources
  • S1In Re: Interplay Between Arbitration Agreements under the Arbitration and Conciliation Act, 1996 and the Indian Stamp Act, 1899 (Supreme Court, 13 December 2023)
  • S2N.N. Global Mercantile Pvt. Ltd. v. Indo Unique Flame Ltd. (Supreme Court, 25 April 2023)
A specialist's draft

An unstamped arbitration agreement is not void; the defect is curable [[S1]].

The contrary view, that such an agreement is unenforceable, was overruled [[S2]].

A later High Court decision has since doubted this [[S9]].

What the reader seesExpanding handles

An unstamped arbitration agreement is not void; the defect is curable [[S1]].

The contrary view, that such an agreement is unenforceable, was overruled [[S2]].

A later High Court decision has since doubted this [[S9]].

04 Statutes

Law as it stands today.

A full database of statutes, Acts and rules, cited from the official text word for word, never from a model's training. A repealed IPC, CrPC or Evidence Act section is read together with the provision it was carried into, and a provision that cannot be found is reported as not retrieved.

  1. A full database of statutes and rules

    Acts, rules and regulations are held in full in our statutes database, not relied on from a model's training.

  2. The official text, word for word

    Every provision is read from the full official text and quoted exactly as it is printed, with a link to the page it came from. Never from what a model remembers.

  3. Watched for amendments

    A pipeline keeps searching India Code and the e-Gazette, so amendments, new rules and notifications reach the database as they are published.

Statutes database: Acts and rules, in full
Arbitration and Conciliation Act, 1996ActFull text
Code of Civil Procedure, 1908ActFull text
Specific Relief Act, 1963ActFull text
Bharatiya Nagarik Suraksha Sanhita, 2023ActFull text
Companies (Accounts) Rules, 2014RulesFull text
Delhi High Court (Original Side) Rules, 2018RulesFull text
05 Court Intelligence

The docket, read every night.

Court Intelligence works on the daily order stream, not just reported judgments, and reaches five court portals live when a question needs a judgment that is not yet in the stream.

Every night

We read the day's orders from the Delhi High Court and the Supreme Court, and our own pipeline processes each one, so the whole docket is ready for your questions by morning.

When you ask

Ask about a bench, a party or a kind of order, and the answer comes from the orders themselves, each linked back to the original.

06 Processing documents

Every page read by the right reader.

Each page takes the cheapest route that can read it properly. A page falls through the cascade only as far as it has to, and nothing is ever guessed.

  1. 1

    Text layer. Kept as it is when it is usable.

  2. 2

    Garbled text check. Catches text layers that look readable but are scrambled.

  3. 3

    OCR. Scanned English pages are read by OCR. Pages it is unsure of go further.

  4. 4

    Script check. A quick check decides whether the page is mainly in an Indian script.

  5. 5

    Indian-script reader. A reader built for Indian scripts, across ten of them.

  6. 6

    Vision reader. Anything still unread is read by a vision model, which writes "illegible" rather than guess.

07 Working with multiple large documents

The right documents, read in full.

The planner picks the documents a question needs. A quick question is answered from the passages that matter. When a question needs the whole matter, such as a timeline or everyone involved, the documents it needs are read in full by readers working in parallel and sharing one context graph, every finding is checked against its document, and the answer is written only from what passed.

  1. 1

    The right documents, read in full. The planner picks the documents the question needs, so not every message reads every document. When the question needs them, readers go through them at the same time, in full, and each finding comes with the exact passage it was taken from.

  2. 2

    One shared context graph. Readers working in parallel write their findings into one context graph, so a date in one document is linked to the same event, party or clause in another, and findings are related to each other rather than just listed.

  3. 3

    Related documents together. A filing goes with its reply, and a run of hearing minutes stays together, so one reader sees each exchange whole. Every document is placed exactly once.

  4. 4

    Every passage found in its document. Each quotation is checked word for word against the document it names. A quotation that is not there takes its finding with it.

  5. 5

    Every date and figure in its passage. A date or amount in a finding must appear in its quotation, or in the sentences beside it. If not, the finding is rewritten to what the passage says, or dropped.

Eight rules every answer passes.

A model can be asked to be careful. These are written in code, and run on every answer.

A closed set of citations. Specialists cite by handle, never by typing a case name. A handle that matches no retrieved source removes the sentence, or the table row, it sits in.

Quotations matched word for word. Against the full text of the source, not a snippet. A passage link is created only from a quotation that matched.

Dates and figures in their quotation. Every date and amount in a finding must appear in the quotation it came from, or the finding is dropped.

Provisions nobody looked up. A sentence that names a provision no specialist actually read is caught and checked.

The court's own copy. Each relied-on case is found on the court's own site and checked for later history.

Binding force from the hierarchy. A decision is called overruled only when a court able to overrule it did.

Links that came back. Any link that did not come back from a court portal or a search is removed.

Misses said out loud. A provision or judgment that could not be found is reported as not retrieved, never replaced with the nearest guess.

What we're building next.

The same principles, carried further: agents that change course mid-run on what the others find.

In design

Behavioural context pool

The shared context graph is already there: every specialist writes what it finds into it. Next, agents read each other's findings while they work. When one agent finds something important, such as an authority that has been overruled, the others change their course of action mid-run instead of finishing on stale ground.

Illustration of one run. Question: does an unstamped agreement bar a Section 11 application?
Statutes researchera1
issue_claimednew_provision

Working

Case-law researchera2
authority_killedissue_claimed

Working

Treatment checkera3
issue_claimed

Working

Document readera4
new_provision

Working

Shared context pool, append-only0 entries
    Verifier

    Reads the whole pool; only it can mark an entry verified

    Synthesiser

    Waits for verified entries

    Your files stay your firm's.

    Private to you and the people you choose, and never used to train AI models.

    Not used for training. Your documents and questions are never used to train AI models, and model request logging is switched off.

    Encrypted. Encrypted in transit and at rest.

    Private inside the firm. Administrators see usage, not content. They cannot open a member's chats or files.

    Searches never carry your files. Web and case-law searches use terms from your question, never the contents of your documents.

    Invite only. Accounts are created by invitation, and every sign-in is logged.

    Deletion is final. Delete a file or chat and it is gone. You can also see and erase anything LegalGita has learned about how you work.

    Learn more about security ›Read the trust centre ›

    See it on your own matter.

    A 30-minute walkthrough with one of your questions, one of your courts, or a set of your documents.

    Book a demoSee pricing